FGR-Grad: rectified full gradient for weakly supervised crack detection

Abstract Detection and segmentation of concrete cracks are critical tasks in structural health monitoring. Although existing fully-supervised learning methods have made significant progress, data acquisition and precise annotation remain costly and challenging, especially in the continuous monitoring of large-scale infrastructure. The high cost and inconsistency of annotations have become limiting factors. To reduce reliance on manual labeling, weakly-supervised semantic segmentation using image-level labels has emerged as an active research focus. However, the pseudo-labels generated by these methods generally exhibit lower quality compared to fully-supervised approaches. To bridge this gap, this study proposes FGR-Grad, a weakly-supervised semantic segmentation method based on the FullGrad algorithm. The approach incorporates a gradient correction strategy (via guided backpropagation (GBP) or rectified backpropagation (RBP) mechanisms) to refine gradient propagation: On one hand, it introduces GBP layer-wise within the FullGrad algorithm, replacing raw gradients with guided gradients to prioritize the propagation of crucial gradient information towards important network parameters, thereby enhancing the model’s focus on key features. On the other hand, it employs the RBP mechanism to prune redundant information in FullGrad saliency maps. This replaces raw gradients with rectified gradients filtered by importance scores, blocking information flow from irrelevant features and eliminating redundancies to highlight the most discriminative target regions. Experimental validation on three public datasets demonstrates that the FGR-Grad algorithm significantly narrows the performance gap between weakly-supervised and fully-supervised methods, with its inference accuracy and visual results outperforming existing weakly-supervised techniques.